Happy Monday!
We want to kick off the week (and our Substack in general) with a series called Skylarking. We’ll be pulling out news, research, and more from around the industry right now and trying to distill what is useful for operators, builders, and tech-industry folks every two weeks. Trying to take the tidal wave of information out there and highlight items along a few angles that are important to our team and the communities and organizations we serve.
TL;DR
Meta says “Me, too!” and releases Muse Code to enter the Agentic Coding Tool space alongside Claude Code, Cursor, Codex, Antigravity, and so many more. At least it doesn’t start with C.
OpenAI released GPT-5.6-Cyber (similar product to Anthropic’s Claude Mythos), Google launched Gemini 3.7 with introductory pricing, and SpaceXAI released Grok 4.6 to compete with Claude Fable/GPT-5.6-Sol. DeepSeek announced price changes (read: increase) alongside their V4-Pro GA release.
Canadian AI darling Cohere has announced partnerships with Universities of Waterloo and Toronto, both focused on AI adoption inside organizations with a view on scaling. The Canadian Government also announced an investment (up to) $240 million for AI compute and platform usage. And yet they still won’t answer our calls.
Ontario Premier Doug Ford announced the Data Centre Playbook in Guelph, Ontario on Friday. They have also announced an ERO looking for public comments about their framework for assessing new data centres in Ontario. With a rise in resistance to new DC builds that will be an interesting read.
The International Joint Conference on Artificial Intelligence (IJCAI) is happening in Bremen, Germany this week! Papers are already listed if you want to get into the weeds on the latest research being discussed and presented.
Shiny & Chrome - What’s New / What’s Useful
The latest and greatest continues to be later and greater(ish). This past couple of weeks have seen quite a few incremental increases to the frontier models across the industry. The most interesting information seems to be less about the overall capabilities and more about what isn’t available and how much what is available changes the pricing. Also, prior to August, the open-weight models were dominating the news cycle so it all feels more like a response to that.
Facebook’s FOMO
Meta has thrown their hat in the ring of agentic coding agents with the release of Muse Code. It’s MacOS/Linux only, WSL2 if you Windows folks need to have a try, and was released for Terminal/CI usage only. No standalone app or IDE extension.
What makes the approach here different compared to the rest of the offerings on the market now is that they are focused on recoverability, not making the most intelligent models. All events, all calls, all tools are logged so that if anything fails/crashes/freezes it can recover and pick up where it left off. It focuses on fan-out approaches using git worktrees to keep things isolated. The real point here is it is meant to be resilient, not efficient. It may take more effort to get to where it needs to get to, but it’ll keep going.
Only thing that caused a head-turn was the wording around the two model IDs announced. The primary model muse-spark-1.2 says it will not use data to improve their products. The Contributor model (muse-spark-1.2-contributor) indicates it not only trains on the code generated, but every prompt, tool call and all the context surrounding coding is now part of training data. It’s significantly cheaper, but basically allows for not only the code, but any other context it pulls in to be part of a training data set. For most organizations that’s Danger Danger.
Cyber Crimes Department
Running a similar playbook that Anthropic did with its Mythos model family, OpenAI has announced a new GPT-5.6-Cyber as part of its Daybreak initiative, harkening to the Project Glasswing initiative that has been ongoing since April of this year.
OpenAI released their own benchmarks and showed the capabilities of their latest model against previous iterations. Without comparing against the rest of the industry, it appears to be a significant increase over their previous models. Most of this won’t be available to the general public, similar to Glasswing, and is meant to be rolled out to Daybreak partners in order to help shore up any Zero Day Exploits the Daybreak-Red model can find. The system card is not yet published, so worth keeping an eye out for when that lands.
Back in a Flash
Not but three weeks after releasing their Gemini 3.6 Flash model, Google’s DeepMind team has released 3.7 Flash. It’s called out quite loudly that this is based on 3.6 Flash, so it shouldn’t be considered a huge leap but rather a tuning of the recently launched model. With that said, it seems to be punching above its weight (pun intended?) compared to its predecessor as it benchmarks above Claude Sonnet, GPT-5.6 Terra and the newly released Muse Spark 1.2 in a few areas it had lagged quite a bit behind.
The one most worth digging at is the AutomationBench by Zapier. It scored 30.4% compared to 3.6’s 17% (and compared to Sonnet/Terra at 10.7/23.6% respectively) which shows a big increase in that coordination and tool usage. This is something that’s been lacking in Gemini’s space from our experience, so this will be top of the tinkering list for the team.
As part of the announcement there was also a Special Introductory Pricing that impacts both 3.6/3.7 of $0.75/1M Input and $3.75/1M Output. That price doubles at the start of 2027. Also worth noting is that the minimal effort has been removed, meaning for folks looking for an upgrade to the cheap 3.5 Flash/Flash-Lite work may need to wait longer still.
The biggest draw for the Gemini Flash models is pure speed. If you haven’t used it for general work, it can feel almost instant in its response rate. The Artificial Analysis analysis shows it hits a great sweet spot on the Smart/Speedy index they use.
New Toys!
Rapid fire on some of the things that made us say “Neat!” when it came across our feeds this past couple of weeks.
Google Sheets Canvas has been announced and it allows you to build out mini-apps within your spreadsheets. It’s available for Business/Enterprise Standard and above, as well as AI Pro/Ultra for individuals. It’d be great to hear from folks who live in Excel/Sheets to see if this is something that would be useful from a BI/Data Visualization standpoint.
Speaking of spreadsheets, Intuit Accountant Suite is set to replace Quickbooks Online Accountant as of next year. Allows for more cross-client insights (according to the announcement) so may be worth trying while it is free for anyone doing books.
Anthropic now allows self-hosted environments for Claude Code (public beta). This will be good for privacy and security wonks who aren’t allowed to sneak their favourite open source solutions into the vendor list. Also Claude in Chrome is now inside Claude Cowork for those who would allow the robots to run their browsers.
AWS open-sourced KiroCrew (which we’ll talk about below) to try and create a governance structure around the concept of OpenClaw. It’s based on their own internal tooling called MeshClaw.
Works on my Machine - Our Team’s Learnings
Normally this section will have a bit of a deep dive into Larkwin’s internal research and usage around AI tools. However, given this is the first issue and we are just starting with a project I’ll tease out what you can expect to see in the next issue.
Don’t mess with our crew
As mentioned above, our team is looking into the implementation of KiroCrew into our stack. With the prevalence of tools like Hemes and OpenClaw, there’s a desire to pull in apps, tools and more to have more autonomous execution. However, most of these are Use At Your Own Risk and really mean it.
KiroCrew seems like a promising alternative, meant to be more isolated but still integrate with relevant tools and apps beyond just using MCP/APIs.
Our team is looking at how this could potentially be extended for Teams. Right now the system is locked to an individual user. Although there is a way to deploy it to a remote-instance, the system for managing team-deploys is a bit clunky. We will be spending time on this problem, trying to see if we can’t make the onboarding of this tool into a distributed team with the same level of security/scalability we would want for any of our Enterprise software.
Stay tuned!
Agentic HR @ Work
Another project in active development within our Labs is a homespun, agentic HRIS. We have had loads of experience (good and bad) with HRIS and ATS systems. Some are fine, but clunky. Some are too limited. Some are too expensive. Finding the right fit is often challenging.
Ahead of onboarding our first team member, we challenged ourselves to automate the process. From interviews to offer letters, from onboarding to mandatory training, from 1on1s to quarterly reviews; our system is meant to help us operate at large, secure scale with our small but mighty team.
We’ll be doing a full write-up on the process in the coming issues.
The Real Heroes - Research and Academia
None of what is technically possible in today’s landscape would be there without researchers spending countless hours determining what is possible, what works, and what doesn’t. Here are a few of the papers that have recently been published in this space that may be worth reading over.
Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents
While working off a small dataset, this study highlights that Agents tend to fail when relying on other skills or agents, making assumptions about their functionality and/or capability.
Does generative AI narrow education-based productivity gaps? Evidence from a randomized experiment
Do AI tools make your best people better or democratize information? This study of 25-45 year-old Argentinians suggests that AI provides more gains for those with lower education (defined as high school diploma only and/or half of a postsecondary degree), and that when the tools are removed those gains are not entirely diminished.
Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks
An interesting study from Imperial College London that shows that Grounding/external knowledge sources can still be susceptible to hallucinations (re: failures). Non-deterministic still is non-deterministic.
The big event to keep an eye on this week will be The International Joint Conference on Artificial Intelligence (IJCAI) in Germany. You can already see the list of papers listed and the various sessions joined to them. Check throughout the week to see more things added.
Standing on Business - Money, Policy & Politics
Profits, Losses, and Bets in Canadian AI
It has been a busy couple of weeks in the country, and definitely in Ontario. One of Canada’s biggest players in the AI space, Cohere, has partnered with both UofT and UW to build out curriculum based on adoption and integrating AI systems into scaling or scaled organizations. It echoes the sentiment from leaders in this space such as Daniel Wigdor (Professor at UofT and founder of AXL) that Canadians lead in the area of Applied Computing and we should be applying that expertise into AI.
Canada has continued to see investment at the Federal and Provincial level. The Deputy Prime Minister announced that they would invest up to $240M into Cohere to scale its AI compute capacity (via CoreWeave, see below). The Federal Government also announced that FedDev would fund $40M into various Ontario organizations to support their growth, with some of those directly mentioning automation/AI systems.
Worth noting re: Cohere is that the $240M is noted as a “partially conditionally repayable contribution” administered by ISED, the largest line item being buying compute from CoreWeave, as reported by The Logic. This means most of this money will be going into the US-based company, rather than a Canadian supplier.
At the Provincial level, Ontario announced their Data Centre Playbook with a view on being transparent on what possible vendors and contractors would need to do in order to be considered by the province. They also have published a call for comments on their assessment framework that closes on September 12th, 2026. Meanwhile, Alberta open-sourced The Velocity White Papers to show their process and mission to make their Government “the most AI-enabled public service in North America.” Which is… a metric to measure against.
Canada has also lost a bit of sovereignty with AMD acquiring Toronto’s inference chip-maker start-up Taalas. While it hasn’t closed, and the release stresses retaining and growing Canadian talent, it is another company that will ultimately have a parent interest elsewhere. KW has also lost Convictional, an e-commerce business that did a hard pivot to be “Slack for the AI era.” Founder Roger Kirkness posted about their journey, the decision and the lessons and information they hope to deliver back to the community.
All this is compounded with two recent StatCan surveys. The first indicating that 19.2% of businesses used AI in the past year (up from 12.2%) but close to half indicated it is not actually relevant to their business. Paired with AI adoption, while showing a 16.8% labour productivity gain, falls closer to 5% when controlled for complimentary capabilities such as training, cloud infrastructure and more.
Two houses, alike in dignity
Anthropic’s recent announcement that they will be imperceptibly watermarking generated text from their Claude models has caused a stir in the industry and wider AI Discourse. They released a more detailed explanation of how it works and noting it is in response to the EU AI Act, alongside other major providers. This highlights that this is likely to be (or to have already been) implemented by other players that operate in the EU (most of them) and isn’t unique to Claude.
On the complete flip side, Google recently enabled users to remove visible watermarks from Gemini’s generated images, videos and music. While in the same week announcing an open source library for adding content provenance information to generated images, text and more. The project, Credentio, is aligned with the Coalition for Content Provenance and Authenticity to help make authorship more transparent. Google has already published a system for verifying AI-generated images, videos, and audio through Gemini Apps.
There continues to be this push-and-pull for the enablement of more and more content creation, while establishing that transparency around that content’s origins is also crucial. Expect this to continue to volley around and be influenced by policy makers in nations worldwide.
B-Sides - Everything else good enough to post
Google Meet now has a button to allow for taking notes for in-person meetings. Folks using competitors like Otter will ask “They didn’t have this already?”
GitHub released a new open framework called Agent Plugins 1.0 that allows you to combine MCP and skills into a single package that can be installed.
DeepSeek’s new pricing includes Off-Peak hours that perfectly align with North American working hours. High five?
This post was not written with any AI assistance and also directly in the CMS like a real hero. Which means you have to deal with all of the improper grammar and weird run-on sentences. Claude is used to help aggregate and factcheck sources as a bit of rubber-ducking. Written posts will always be marked on the bottom with how/if AI is used.